Turns out that Wrnodedev writes .ic statements, which only work for seeding the initial operating point of another transient analysis. What I can't find for ngspice's AC (or NOISE) analysis is to to tell it to use a precomputed operating point. In spectre, you can do this with ac readns="foo.txt". Definitely not an original Berkeley spice feature, though.

@Boris Murmann Also consider using wrnodev at a time point where the circuit is settled to a reasonably stable state, so not immediately after a clock or reset edge or any other signal change.

Here's a test case: Take a voltage source, a switch and a capacitor with a large resistance across. Run a transient sim with the switch on to charge the capacitor and then off to let that voltage sit and save the state shortly after. Now run OP/AC/NOISE and see if that voltage is still properly initialized in these analyses' op point. Spectre does this correctly using the readns option. The parallel resistance discharges the cap under DC conditions. That's why there should not be any DC analysis done in this case; the simulator must be told to just leave that initial voltage as is. That is what readns does in spectre. In other words, an op analysis that is given some hints does not preserve the state, it will just converge to a valid DC solution, which has the cap fully discharged. This is really nothing exotic, but bread and butter functionality for anyone doing switched capacitor circuit design involving amplifiers, dynamic common mode feedback loops, etc. These circuits never operate around bias points for t->infinity, they operate at bias points that are dynamically defined in their clock cycles.

Commenting out the optran statement gives 0V, while having it gives the desired 1V. 😀

Of course, as Stefan pointed out, this has be used with care, and you have to "know what you're doing."